acceptodds
Withdrawn from ICLR 2027

Stable-Structure-Guided Spectral Corruption for Diffusion-Based Multivariate Time Series Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion-based reconstruction has shown promising results in multivariate time series anomaly detection, using discrepancies between observations and their reconstructions to identify anomalies. However, isotropic corruption may affect the input information available to the denoiser, raising the question of how to design anisotropic corruption based on structural characteristics. Existing frequency-aware diffusion methods typically allocate corruption under prior assumptions that equate absolute frequency location or spectral energy with information value; these assumptions may overlook normal high-frequency transients and low-energy frequency components that encode reproducible, stable structures. To address this issue, we propose a stable-structure-guided diffusion framework. The framework constructs adaptive soft frequency bands from local window features and combines local spectral prototype support, cross-window reproducibility, and cross-variable relational consistency provided by a stable reference bank to form band-level structural evidence. This evidence guides the allocation of corruption strengths to preserve components supported by stable patterns during denoising. Across five multivariate benchmark datasets, the framework improves mean F1 by 33.10% over the state-of-the-art diffusion model and by 24.86% over the state-of-the-art model among all evaluated baselines, which comprise 41 representative anomaly detection methods.

Primary area:
generative models
Keywords:
time series anomaly detection, diffusion models, frequency-domain modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ojpYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates multivariate time series anomaly detection using diffusion models and proposes a stable-structure-guided spectral corruption framework. Unlike existing methods that allocate noise primarily based on frequency location or spectral energy, the proposed method constructs a stable structure reference bank by combining adaptive soft frequency bands, local spectral prototypes, cross-window reproducibility, and cross-variable relational consistency. The resulting structural evidence is then used to adjust the noise strength across frequency bands. Experiments on five multivariate time series benchmarks show competitive anomaly detection performance compared with existing methods.

Critical strengths

1.This paper introduces stable structural consistency as a criterion for spectral noise allocation, rather than relying solely on frequency location or spectral energy. This provides an interesting perspective on frequency-aware diffusion models.

2.This paper includes extensive comparisons on five multivariate time series benchmarks. The ablation studies and case studies further support the effectiveness of the proposed method, which achieves competitive performance against existing approaches.

Critical weaknesses

1.The method combines several structural features to estimate stability and adjusts the noise strength accordingly. However, the rationale for assigning weaker corruption to frequency bands with higher structural consistency is not sufficiently justified.

2.The proposed framework introduces additional operations (e.g., frequency-band generation and clustering), but provides limited analysis of the overall training and inference costs.

3.Some baseline results are taken from a unified evaluation framework. It remains unclear whether the preprocessing, threshold selection, and experimental settings are fully consistent across all compared methods.

4.Figure 2 contains too many operations and details, making it difficult to identify the main contributions of the proposed framework. I suggest simplifying the figure and highlighting the key innovations more clearly.

5.The complementary roles of these structural features are mainly supported by ablation results, with limited analysis of how they interact.

6.The manuscript could be further improved in terms of presentation and writing. For example, some equations are missing appropriate punctuation, such as commas at the end of equations.

Questions for authors

1.Could the authors further explain the relationship between stable structural evidence and noise allocation? In particular, why should frequency bands with higher structural consistency receive weaker corruption? Are there any types of anomalies or data characteristics for which this assumption may not hold?

2.Could the authors provide a comparison of training time, inference time, and computational overhead against the main diffusion-based baselines under the same experimental settings? This would help assess the additional cost introduced by the structural modeling components.

1 comment
IDReplyDetails
W1,Q1Weak noise is applied to high-support bands to preserve normal structure, while strong noise is applied to low-support bands to avoid learning anomalous components (Details 1).Details 1: The goal of anomaly detection is to reconstruct the normal trajectory rather than the original data. The original data may contain background noise or anomalies. Applying different noise levels according to support is a weighting choice: weak noise preserves the normal structure needed for learning, while strong noise reduces the risk of learning potentially anomalous components, making reconstruction rely more on normal patterns.
W5On page 9 of the paper,Figure 3 illustrates how structural features work together to identify anomalies (Details 1).Details 1: We have uploaded high-resolution supplementary images to the https://anonymous.4open.science/r/Re-9CB5/. We will include these images as Figure 4 in the revised manuscript to provide additional context for Figure 3.Figure 3a: The normal trajectory is constant (see contextual images), and the anomaly appears as an abrupt change. The high-frequency components of this change receive low support from the reference bank and are strongly corrupted with noise. The model ultimately generates the constant structure of the normal trajectory and successfully detects the anomaly. MATSD and FSDI are affected by the abrupt anomaly, producing reconstructions that follow the anomalous fluctuations.Figure 3b: MATSD reconstructs the anomaly because of its low-frequency variation, while FSDI reconstructs it because of its high-spectral-energy features. The reference bank assigns low support to the abrupt change, leading to strong noise corruption. The model reconstructs the normal constant structure and identifies the anomaly. MATSD reconstructs most constant intervals well but still preserves anomalous changes within the anomalous intervals.Mechanism: The frequency-band generator separates the high-frequency components of abrupt anomalies from the low-frequency components of the stable constant pattern. The reference bank scores each band according to its distance from the stable structure at the cluster center, assigning low support to bands far from the center. The abrupt anomalies in Figure 3 all receive low support because of these deviations. The noising module applies strong noise to these components, causing them to be corrupted into random Gaussian noise early in the diffusion process. During denoising, the model does not generate these high-frequency abrupt anomalies. The resulting constant reconstruction produces a large reconstruction error, identifying the anomalous portion.
W2,Q2On SWaT (944,919 × 51, RTX 5090, FP32), the time/epoch:· Pretraining: 28.59 minutes.Reference-bank construction: 3.26 minutes.Main training: 23.84 minutes.Inference: 4.4 minutes.Memory overhead: the reference bank uses 192.60 MiB CPU memory(Details 1).Details 1: Pretraining additionally learns the reference bank, making it more expensive than main training. We predict the original data directly rather than using DDPM’s iterative sampling, avoiding the multi-step inference overhead.
W3We follow TAB’s parameter configurations (e.g., window_size=100) and use overlapping windows at inference, as is common in recent work. (Details 1).Details 1: We successfully reproduced 55.7% of the tasks, with metric differences within 1.33%. Additionally, TAB uses non-overlapping windows to reduce computation, whereas we use overlapping windows for all inference, which is common practice in recent work.
W4,W6Thank you for the suggestions. We will simplify and redraw the figures , and improve the architecture diagram and formatting throughout the manuscript.
Reviewer_dBTXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper highlights that quantifying discrepancies in diffusion-based reconstruction has shown promising results for time series anomaly detection. However, existing frequency-aware diffusion methods may overlook normal high-frequency transients and low-energy frequency components that capture reproducible and stable structures. To address this limitation, the authors propose a stable-structure-guided diffusion framework that introduces diffusion corruption at different levels across frequency bands.

Critical strengths

• The authors clearly illustrate the problem by demonstrating how removing different frequency components affects time series reconstruction.

Critical weaknesses

• The paper incorporates multiple modules, which is commendable. However, the rationale for introducing each module and its contribution to the overall framework could be explained more clearly. E.g., the authors could provide an overview at the beginning of the methodology section to explain the overall framework, the motivation behind each module, and how the modules work together. • It is unclear which modules represent novel contributions, and which are adopted from prior works. Furthermore, the methodology introduces numerous functions and mathematical symbols without sufficient explanation, making it difficult to follow. For example, in Eq. (12), the authors should clarify what zm,djz_{m,d}^{j} represents and which features are being clustered.

Questions for authors

• Eq. 10 contains multiple functions. Could the authors define these functions to improve clarity and facilitate understanding? • Is Eq. (15) intended to represent the denoising network? The authors state that it follows CSDI. However, CSDI trains its denoising network to predict the noise added during the diffusion process rather than directly reconstructing the original data. The authors should clarify the function of Eq. (15) and explain how their implementation differs from or follows the original CSDI formulation. • Appendix C presents several equations, such as Eq. (18), whose derivations are unclear. Could the authors elaborate on how these equations are derived and explain the motivation behind their formulation? • • In Eq. (12), e_m is associated with three deviation components. However, the formulations of reproducibility deviation D^{rep} and relational deviation D^{rel} appear identical, while prototype deviation D^{proto} differs. Could the authors clarify the rationale behind these formulations and explain why only D^{proto} is defined differently?

1 comment

Thank you for your recognition and guidance.

IDReplyDetails
W1Thank you for the suggestion. We will add Section 4.1 (Framework Overview) to the revised manuscript.
W2W2-1: Thank you for the suggestion. Our novel contribution is a structure-guided noising strategy. To implement it, we design a Gaussian-kernel frequency-band generator and a reference bank that guides the allocation of diffusion noise.W2-2: The mathematical details are given in Appendix C, and we will improve their presentation in the revised version. For example, the notation in Eq. (12) is explained in Details 1.Details 1: Notation in Eq. (12)· zm,djz_{m,d}^{j}: The dd-th component of the jj-th feature group in the mm-th frequency band of the current window.· mm: Frequency-band index.· jj: Feature group (proto, rep, or rel). Each group assesses structural normality from one perspective. The three groups jointly represent structural support with equal weights.· dd: Feature coordinate. proto has 3 dimensions, rep has 1 dimension, and rel has 8 dimensions.
Q1Eq. (10) compresses the inter-channel relation matrix. Computing pairwise relations for CC channels yields a C×CC \times C matrix, which is computationally expensive to process directly. Eqs. (20)–(22) in Appendix C explain the compression.· AmA_m: Symmetric C×CC \times C matrix of normalized cross-spectral relations between channel pairs in frequency band mm — Eq. (20).· qmq_m: C(C−1)/2C(C-1)/2-dimensional vector of the entries above the diagonal of AmA_m, extracted by vec_upper — Eq. (21).· rmr_m: 8-dimensional vector of summary statistics computed from qmq_m by Stats — Eq. (22).· rr: 8M8M-dimensional representation of the current window, obtained by concatenating the rmr_m vectors from all MM bands — Eq. (10).
Q2Q2-1, Q2-3: Eq. (15) specifies the inputs and output of our denoising network. We retain the CSDI architecture with different conditioning inputs (see Details 1) and use a single forward pass at inference instead of iterative sampling.Q2-2: We predict the clean window rather than the added noise and use that prediction as the final reconstruction to avoid the cost of iterative DDPM sampling (Details 2).Details 1: Eq. (15) takes the corrupted window XnX_n at diffusion step nn and two conditioning inputs (lines 348–350): CnmeanC_n^{\mathrm{mean}} is the mean corruption strength across frequency bins, and CnspecC_n^{\mathrm{spec}} gives the deviation from that mean at each frequency bin. We follow CSDI's conditional-input design but condition on corruption levels rather than observed values.Details 2: With a 50-step forward diffusion schedule, standard DDPM sampling requires 50 denoiser evaluations per window. DDIM can skip intermediate steps, but a 12-step sampler still requires 12 evaluations. As an engineering optimization, we reconstruct the clean window with a single denoiser evaluation.
Q3Appendix C expands the feature computations in the method section. These definitions come from information theory, digital signal processing, Fourier analysis, pattern recognition, and probability theory (Details 1).Details 1:· Eqs. (17) and (18): Spectral energy, spectral entropy, and spectral flatness quantify the power distribution and randomness of a frequency band.· Eq. (20): Normalized cross-spectrum measures the phase-coherent relationship between channel pairs.· Eqs. (24) and (27): Gaussian smoothing operations reduce local variability in spectral features.
Q4The three components are computed in different feature spaces and converted into unified support via exponential mapping. proto introduces the temperature τ\tau to mitigate the influence of spectral peak energy extrema (Details 1).Details 1: rep and rel measure “distance from cluster center,” but weighted fusion requires “similarity to cluster center.” The exponential mapping e−D/τe^{-D/\tau} converts deviation DD into support on a common scale. Spectral peak strength varies across windows, so local extrema can dominate the distribution. proto introduces the temperature τ\tau to mitigate this.
Reviewer_p743OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper allocates spectral diffusion noise according to support from normal structures. Adaptive soft frequency bands and a reference bank describe spectral prototypes, cross-window deviations, and cross-variable relations. Strongly supported bands receive weaker corruption. A conditional diffusion model reconstructs the sequence, and residuals provide anomaly scores.

Critical strengths
  1. Useful normal information can occur outside low-frequency or high-energy components. Protecting information based on agreement with normal operating structures is therefore a meaningful alternative to fixed spectral heuristics.

  2. The results demonstrate potential without relying on PA. The reported mean F1 is approximately 42.15 versus 31.67 for PhysDiff, consistent with the stated 33.10% relative improvement. Multiple metrics and component ablations support the usefulness of the full implementation, although performance varies by dataset.

Critical weaknesses
  1. The experiments do not isolate the benefit of corruption allocation from the reference bank’s discriminative information. The reference deviations already define a potential anomaly detector, but direct scoring with them is not evaluated. Since frequency-selective diffusion has been studied in MATSD and FSDI, the new contribution depends on demonstrating the value of structural guidance. Reference-only scoring and comparisons with matched backbone, conditioning, and noise energy are needed to establish that attribution.

  2. Figure 2 states that the band generator is reinitialized, whereas Appendix C.3 says it is retained and frozen. This changes its compatibility with the reference bank. Moreover, Eqs. (26)–(28) preserve an unweighted average over one-sided rFFT bins, which does not by itself preserve time-domain total noise energy: interior bins and DC/Nyquist have different multiplicities. The intended budget and actual normalization need clarification.

Questions for authors
  1. Which training procedure produced the reported results: reinitializing the band generator or retaining and freezing it? How is compatibility with the frozen reference statistics maintained?

  2. Is the conserved quantity the unweighted one-sided spectral variance or expected time-domain noise energy? How does the implementation handle Parseval multiplicities, and which quantity is matched across baselines?

  3. How well do the reference-bank deviations or fused support detect anomalies without diffusion reconstruction? Which matched control separates the benefit of support-guided corruption from the bank’s existing discriminative information?

2 comments

Thank you for recognizing the value of protecting information based on normal operating structures and for the detailed evaluation of our experimental results.

IDReplyDetails
W1,Q3W1-1,Q3-2: The reference bank estimates structural support for each frequency band (high support → weaker noise; low support → stronger noise), guiding the diffusion process to preserve normal patterns (Details 1).Q3-1: Table 1 compares detection performance across ablated configurations, including reference-bank-only detection.W1-2: Table 2 demonstrates the gains from structure-guided noising by replacing MATSD and FSDI strategies with our approach while keeping their backbones unchanged (Details 2).Details 1: The reference bank constructs a clustering-based representation of structural normality. Table 1 shows F1 scores for K-means clustering, reference-bank-only detection, DDPM, full-reference-bank, and full. K-means achieves strong recall (0.832) on PSM , but this advantage diminishes under point-adjusted evaluation.Details 2: Lines 50–89 of the manuscript explain the limitations of MATSD and FSDI.Both strategies rely on fixed spectral heuristics. Our structure-guided allocation adapts to the actual normal operating structures in each dataset. Table 2 reports the relative F1 improvement (%) after replacing their noising strategies with ours.
W2-1,Q1The band generator is initialized for main training, while the reference bank is frozen after pretraining. This two-stage design resolves the mutual dependency between band generation and reference statistics (Details 1).Details 1: We use John Wheeler's description of general relativity—"Spacetime tells matter how to move; matter tells spacetime how to curve"—to illustrate the dependency. The features used by the reference bank depend on the frequency-band samples produced by the generator, while updates to the generator depend on the reconstruction error obtained after reference-bank-guided noising. If both are updated simultaneously from initialization, different random seeds lead to different cluster formations, destabilizing the reference statistics.Drawing on the Banker's algorithm for deadlock resolution, we handle the initialization dependency during pretraining: the band generator and reference bank co-evolve until convergence, then the reference statistics are frozen. During main training, the generator adapts to reconstruction objectives while the frozen reference bank provides consistent structural guidance.
W2,Q2Q2-1, W2-2: The conserved quantity is the unweighted mean noise variance across one-sided rFFT bins, not the expected time-domain noise energy (Details 1).Q2-2: All bins, including DC and Nyquist, receive a weight of 1. With torch.float32, the relative error in time-domain noise energy is approximately 0% (Details 2).Q2-3: At diffusion step nn, all compared methods use the same global noise-variance coefficient as DDPM, 1−αˉn1-\bar{\alpha}_n.Details 1: We sample real-valued Gaussian noise in the time domain and use rfft and irfft to apply variance scaling by frequency band. Let gkg_k denote the variance multiplier at bin kk. Eqs. (26)–(28) normalize the multipliers so that their unweighted mean is 1. We will make this explicit in the revised manuscript submitted for review.Details 2: For an even transform length LL, Parseval's theorem gives a relative deviation of (2−g0−gNyq)/L(2-g_0-g_{\mathrm{Nyq}})/L from the expected energy of unscaled noise. Only the DC and Nyquist multipliers contribute, with a scaling factor of 1/L1/L.

Table 1: F1 comparison showing the contribution of diffusion reconstruction over reference-bank-only detection

DatasetK-meansBank-onlyDDPMFull-only-bankFull
MSL18.4419.2216.2728.7530.45
SMAP6.3016.4815.3220.1425.51
PSM57.6033.6925.9031.0340.93
PSM (PA)61.1973.56----94.61
SWaT30.7459.7750.0461.8168.78
SMD16.9030.0019.4341.1945.10

Table 2: Relative F1 improvement (%) after replacing MATSD/FSDI noising strategies with structure-guided allocation

DatasetMATSDFSDI
MSL+30.96%+27.44%
SMAP+92.69%+38.12%
PSM+7.07%+65.11%
SWaT+136.35%+243.50%
SMD+0.21%+251.79%

Thanks for the clarifications and the additional experiments. The reference-bank-only results and the comparisons using the MATSD/FSDI backbones are helpful. But some of my concerns remain, so I’m keeping my rating unchanged for now:

  1. The ''full-only-bank'' experimental results are consistent with the original S✗/E✓/B✗ configuration, which also employs guided corruption. This setup requires a more precise definition. I think a matched comparison should keep the backbone, bank information, conditioning, and noise energy fixed while changing only the noise allocation.

  2. Sec. 4.4 introduces 50-step deterministic DDIM sampling, while the response to reviewer #dBTX describes a single-step denoising forward process that does not involve iterative sampling. It is essential to explicitly state which setting was used for the reported results. Providing the code and configs for these experiments will help clarify the specifics of each variant and facilitate the reproducibility.

  3. Fig. 2 indicates that the ''band generator'' is re-initialized, whereas Appendix C.3 states that it is retained and kept in a frozen state. And the response mentions that it continues to undergo adaptive adjustments during the main training phase. Please clarify which specific procedure was used to obtain the reported results.

  4. If the generator does indeed change, how is it ensured that its output features remain compatible with the frozen band bank?

If these issues can be clearly explained, I would be willing to increase the score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.